Comprehensive Cancer Center (CCC Munich LMU), LMU University Hospital Munich, Pettenkoferstraße 8a, Munich, Germany.
Roche Pharma AG, Grenzach-Wyhlen, Germany.
Clin Exp Med. 2024 Apr 10;24(1):73. doi: 10.1007/s10238-024-01336-w.
Personalized medicine offers targeted therapy options for cancer treatment. However, the decision whether to include a patient into next-generation sequencing (NGS) testing is not standardized. This may result in some patients receiving unnecessary testing while others who could benefit from it are not tested. Typically, patients who have exhausted conventional treatment options are of interest for consideration in molecularly targeted therapy. To assist clinicians in decision-making, we developed a decision support tool using routine data from a precision oncology program.
We trained a machine learning model on clinical data to determine whether molecular profiling should be performed for a patient. To validate the model, the model's predictions were compared with decisions made by a molecular tumor board (MTB) using multiple patient case vignettes with their characteristics.
The prediction model included 440 patients with molecular profiling and 13,587 patients without testing. High area under the curve (AUC) scores indicated the importance of engineered features in deciding on molecular profiling. Patient age, physical condition, tumor type, metastases, and previous therapies were the most important features. During the validation MTB experts made the same decision of recommending a patient for molecular profiling only in 10 out of 15 of their previous cases but there was agreement between the experts and the model in 9 out of 15 cases.
Based on a historical cohort, our predictive model has the potential to assist clinicians in deciding whether to perform molecular profiling.
个性化医学为癌症治疗提供了靶向治疗选择。然而，是否将患者纳入下一代测序（NGS）检测的决策尚未标准化。这可能导致一些患者接受不必要的检测，而其他可能受益的患者则未接受检测。通常，已经用尽常规治疗选择的患者是考虑进行分子靶向治疗的对象。为了协助临床医生做出决策，我们使用精准肿瘤学计划中的常规数据开发了一个决策支持工具。
我们使用机器学习模型对临床数据进行训练，以确定是否应针对患者进行分子分析。为了验证该模型，我们使用多个具有特征的患者案例小插图将模型的预测结果与分子肿瘤委员会（MTB）的决策进行了比较。
预测模型包括 440 名接受分子分析的患者和 13587 名未接受测试的患者。高曲线下面积（AUC）分数表明，在决定进行分子分析时，工程特征很重要。患者的年龄、身体状况、肿瘤类型、转移和先前的治疗是最重要的特征。在验证期间，MTB 专家在其之前的 15 个案例中的 10 个案例中做出了同样建议患者进行分子分析的决策，但专家和模型在 15 个案例中的 9 个案例中达成了一致。
基于历史队列，我们的预测模型有可能帮助临床医生决定是否进行分子分析。
